Anthropic Weakened Safety Filters, Signed an AI Cyberattack Warning Letter, Then Shipped Mythos 5.1 Anyway
AgentBlackVeil · reddit · 2026-09-03
A Reddit post lays out an awkward timeline: on Aug 21 Anthropic published "bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders" — in practice loosening filters, with the Claude Code cyber filter triggering 60% less often and the biology filter down 85% on standard medical questions. Six days later, on Aug 27, Anthropic co-signed a letter with 115 other companies (including Google and OpenAI) calling AI cyberattacks an "imminent threat." Then on Sept 1 it shipped Mythos 5.1 anyway — the same model the US put export controls on in June, forcing Anthropic to disable Table 5 until the ban lifted June 30.
The author concedes counterpoints: the percentage drops are false-positive reductions, not guardrail removal; the model still refuses to write exploits, and access is heavily gated and US-only — but the vetting is done by Anthropic itself. The real question isn't whether loosening is defensible, but who decides the timing of these releases.
More from Models
- MOSS SFT vs Flux 3 Klein vs SD 3.0: Same-Prompt Comparison Exposes SD3's Age — ZootAllures9111 · 2026-10-18
- Quick take: GPT-6 matches Opus 5.5 at coding, far ahead at computer use — _vztu · 2026-10-18
- Open models now above Sol-level, outpacing expectations as community awaits V4.1 Pro — yacineMTB · 2026-10-18
- Codex took 3 days, Opus 5.5 nailed it in an afternoon — banned user's video protest — dotey · 2026-10-18
- Yacine MTB: Users will remember how Anthropic treated them once intelligence proliferates — yacineMTB · 2026-10-18
- Mistral's Timestral forecasting models start at 5M params and run on plain CPUs — lmoroney · 2026-10-18